Pith. sign in

REVIEW 3 major objections 5 minor 21 references

The paper argues that vision-language and vision-only driving policies converge into a shared decision space but keep complementary long-tail strengths, so selecting the better trajectory per scenario raises the driving score substantially

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-04 06:02 UTC pith:DDKKFTY6

load-bearing objection The empirical complementarity between these two specific trained policies is real and worth referee time, but the paper's central 'VLM vs vision-only' attribution is confounded by a training-protocol mismatch. the 3 major comments →

arxiv 2602.10719 v2 pith:DDKKFTY6 submitted 2026-02-11 cs.RO cs.CV

From Representational Complementarity to Dual Systems: Synergizing VLM and Vision-Only Backbones for End-to-End Driving

Sining Ang , Yuguang Yang , Chenxu Dang , Canyu Chen , Cheng Chi , Haiyan Liu , Xuanyao Mao , Jason Bao
show 3 more authors
Xuliang Bingchuan Sun Yan Wang
This is my paper
classification cs.RO cs.CV
keywords end-to-end drivingvision-language modelsvision-only backbonesrepresentational similarityCKAsparse autoencodertrajectory selectionNAVSIM
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper studies whether vision-language models (VLMs) and standard vision-only encoders in end-to-end driving are redundant or complementary. It claims that after policy learning, the two branches share a large decision-relevant subspace, yet each retains a small, behaviorally meaningful residual subspace: VLMs become more assertive in long-tail semantic and interaction-heavy scenarios, while vision-only encoders stay more conservative in geometry-dominated cases. This complementarity is exploitable—a per-scenario oracle selecting the better trajectory raises the NAVSIM PDMS from 90.80 to 93.58, a learned trajectory scorer (HybridDriveVLA) reaches 92.10, and a fast-slow variant (DualDriveVLA) that invokes the VLM in only ~15% of scenarios achieves 91.00 with a large throughput gain. The paper's central message is that the VLM-versus-vision-only choice is not a simple accuracy–cost trade-off but a division of labor that can be harvested at the trajectory level.

Core claim

Under a unified VLM-hidden + diffusion-policy paradigm, the paper shows that VLM and vision-only branches share a substantial common decision subspace after policy learning—linear CKA rises from about 0.22 at the backbone level to about 0.54 at the planner level, and the highly aligned CCA subspace expands—but the two branches still retain non-transferable residual subspaces. Using a Shared-Unique Sparse Autoencoder, the paper demonstrates that these residuals are behaviorally relevant: cross shared-only reconstruction is worse than self-reconstruction (nonzero self–cross gap), and the residual differences translate into systematic driving-style divergence. Specifically, the VLM branch wins

What carries the argument

The central tool is the Shared-Unique Sparse Autoencoder (SAE), which decomposes each branch's features into a shared latent (d_s = 64) and a branch-unique latent (d_u = 16) with an additive linear decoder, trained under full reconstruction, shared-only self and cross reconstruction (the cross term enforces interchangeability), VICReg-style anti-collapse, orthogonality, and sparsity. Its primary metric is the self–cross gap Δ_cross = R²_self − R²_cross, measuring how much shared factors fail to transfer across branches. Complementing this are linear CKA and PCA-whitened CCA for alignment analysis, and a trajectory-level scorer (inspired by a scoring module that predicts PDMS sub-components)

Load-bearing premise

The paper assumes the observed behavioral difference between the two branches is caused by the vision-language nature of the VLM backbone, but the VLM branch also received a language-domain-adaptation stage and has 2 billion parameters, so scale or training protocol could be the true cause.

What would settle it

Train a vision-only backbone with the same scale (about 2B parameters) and give it an identical domain-adaptation pretraining stage on driving video-text data without any language supervision; if the near-symmetric decisive-win tails (about 2–3% per side) and the best-of-2 oracle gain (now +2.78 PDMS over the VLM baseline) shrink sharply or disappear, then the complementarity is not specific to vision-language models.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Replacing an expensive VLM with a cheap vision-only backbone costs only a moderate average score drop (90.80 → 88.88) because most decision factors are shared; the cost is paid in long-tail scenarios where the vision-only policy fails.
  • A per-scenario oracle that picks the better of the VLM and ViT trajectories raises PDMS by +2.78 (90.80 → 93.58), establishing a clear upper bound from complementarity.
  • A lightweight learned trajectory scorer on a structured 11-candidate set (endpoints plus interpolations) captures most of that bound (92.10, +1.30) without retraining the policies.
  • DualDriveVLA, which runs the vision-only fast path by default and invokes the VLM only when scorer confidence is low, achieves 91.00 PDMS (+0.20) with about 3.2× throughput gain and roughly 85% of scenarios on the fast path.
  • Representation-only gating—using CKA, CCA, or SAE energies—fails to reliably predict per-scenario winners, indicating that complementarity is a trajectory-level phenomenon and motivating scorer-based selection.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The paper's complementarity evidence is built on exactly one VLM instance (InternVL-2B) and primarily one vision-only instance (ViT-large); if the 2B scale or the extra language-domain-adaptation stage, rather than language semantics, drives the assertive long-tail behavior, the 'vision-language' attribution would need revision, though the fast-slow selection system would still function.
  • The near-symmetric decisive-win tails (about 2–3% of scenarios for each side) hint at a general 'division of labor' principle: any two sufficiently different backbones may exhibit similar complementarity, making ensemble selection with a scorer a cheap way to improve E2E driving without more data or larger models.
  • The threshold sweep in DualDriveVLA shows a smooth accuracy-compute trade-off, suggesting the VLM invocation rate could be tuned at deployment to meet a real-time latency budget, effectively turning the method into a continuous resource controller.
  • The SAE's unique residual factors, which the paper shows encode meaningful behavioral differences, could serve as auxiliary signals for scenario-level risk or uncertainty estimation, since static confidence scores alone do not capture which branch will win.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper studies, within a unified frozen-VLM-hidden + diffusion-policy planning architecture (RecogDrive), how a VLM visual backbone (InternVL-2B) differs from standard vision-only backbones (ViT, ResNet, EVA-CLIP). It asks three questions: whether representations become more aligned after policy learning (RQ1), whether residual differences translate into scenario-level behavioral complementarity (RQ2), and whether this complementarity can be exploited by a trajectory scorer and a fast-slow deployment (RQ3). The main empirical claims are: (i) decision-level features are far more similar across backbones than backbone-level features (linear CKA rises from ~0.22 to ~0.54); (ii) VLM and ViT policies are complementary on a long tail, with each decisively winning on about 2–3% of scenarios and an oracle best-of-2 reaching 93.58 PDMS versus 90.80 for the VLM baseline; (iii) a learned trajectory scorer (HybridDriveVLA) reaches 92.10 PDMS, and a DualDriveVLA variant calling the VLM only 15% of the time reaches 91.00 PDMS with a reported speedup. The paper concludes that VLM and vision-only branches retain non-transferable, behaviorally relevant residual subspaces and that these reflect systematic VLM-versus-vision complementarity.

Significance. If the causal interpretation were established, this would be a valuable analysis-to-system contribution. The paper provides direct evidence that the two specific trained policies—the frozen InternVL-based branch and the joint-trained ViT branch—are complementary at the scenario level: the oracle best-of-2 gain and the conservative win counts do not rely on the learned scorer and therefore independently support set-level complementarity. The paper also honestly documents a negative result (representation-only gating fails), which strengthens the later move to trajectory-level selection. However, the central conceptual claim—that this complementarity is attributable to 'VLM versus vision-only'—is not supported by the experiments as described, because the branches differ not only in backbone type but also in training protocol and parameter scale. The measured complementarity remains interesting even without the causal attribution, but the paper's framing, title, and conclusions promise more.

major comments (3)
  1. [§3.2 and Appendix A.1.2] The fairness claim in §3.2 ('differences are restricted to the backbone initialization and which backbone parameters are updated') is contradicted by Appendix A.1.2. The VLM branch first undergoes an image–text domain-adaptation stage with trajectory supervision and is frozen during imitation learning and RL; the vision-only branches skip domain adaptation and are trained jointly with the planner during IL, then frozen only for RL. The branches therefore differ in at least three entangled ways: backbone architecture, parameter scale (2B vs ~300M), and training protocol. Consequently, the RQ1 shared/unique-subspace results, the RQ2 win counts, and the RQ3 oracle upper bound all measure complementarity between these two particular protocol-instances. They do not establish that complementarity is caused by 'VLM-ness' rather than by the domain-adaptation stage, the frozen-vs-trainable distin
  2. [§5.1 and Appendix C] The decisive-win counts (257 vs 253 at τ=0.2, 159 vs 153 at τ=0.5) are central evidence for complementarity, but the counting protocol is under-specified. Appendix C defines a per-seed significant win as |Δ_{r,i}|>τ, yet the main text reports only aggregate counts without stating how the three seeds are combined (e.g., majority vote, intersection, or union) or whether the reported counts are from a single seed. If these are single-seed numbers, the 'stability-aware' claim is weakened; if multi-seed, the aggregation rule must be stated. Please report per-seed counts and the aggregation rule.
  3. [Tables 4–5 and §6] The headline system gains—92.10 for HybridDriveVLA (+1.30) and 91.00 for DualDriveVLA (+0.20)—are reported without any error bars, number of seeds, or variance information. Given the small margin of the DualDriveVLA improvement, this is not a presentation detail: a single-seed run with realistic closed-loop simulator noise could change the conclusion. At minimum, report mean±std over multiple seeds for the main baselines and proposed methods, and state whether the same evaluation protocol as in Appendix C was used.
minor comments (5)
  1. [Abstract vs. §6.2/Conclusion] The throughput speedup is inconsistently reported: the paper-level abstract says 'about a 1.9× latency speedup' while the full-text abstract, §6.2, and the conclusion say '3.2×'. Please reconcile.
  2. [Table 7] Table 7 contains placeholder text ('Fill with the same evaluation metrics used in Table 2') rather than actual values. This must be completed before publication.
  3. [§6.1, Eq. (5)] The scorer's training set is not clearly specified. If D is the navtest set or overlaps with the evaluation scenarios, the selection result would be circular. Please state explicitly that the scorer is trained only on training-split scenarios and never sees navtest labels.
  4. [§5.2] The driving-style analysis (66% faster VLM, etc.) is reported without a quantitative measure of dispersion or a significance test. Since the styles are claimed to be systematic, please provide confidence intervals or a paired test.
  5. [Figure 4] The right y-axis is labeled 'throughput / latency', which is ambiguous. Clarify whether it is throughput (samples/s) or latency ratio.

Circularity Check

0 steps flagged

No significant circularity: central results are benchmarked against official NAVSIM metrics and the scorer is a surrogate, not the evaluation itself.

full rationale

The paper's central findings—representational convergence (CKA/CCA/SAE), long-tail complementarity (win counts, oracle 93.58), and system gains (92.10 and 91.00)—are all evaluated by the official NAVSIM PDMS/EPDMS metrics, not by the paper's own fitted quantities. The Shared–Unique SAE is used diagnostically, and the paper explicitly reports that representation-only gating fails (90.96 vs 90.80 baseline, far below the 93.58 oracle), which is a negative result rather than a circular validation. HybridDriveVLA trains a trajectory scorer to predict PDMS sub-components (Eq. 5) and selects trajectories via the predicted meta-score (Eq. 6), but the reported PDMS is then computed by the official NAVSIM evaluator; this is a standard surrogate-optimization pipeline, not a case where the predicted score is presented as the measured outcome. RQ2's win counting uses per-scenario PDMS and the oracle best-of-two is explicitly a definitional upper bound, not a derivation. The Appendix A.1.2 protocol difference—VLM branch undergoes domain adaptation and is frozen, while vision-only branches are trained jointly during IL—is a real attribution confound for the causal 'VLM vs vision-only' interpretation, but it is a validity limitation, not circularity: no equation reduces to its own input, and no fitted parameter is renamed as a prediction. RecogDrive is used as a base implementation and DrivoR as an external scorer; no load-bearing self-citation chain or imported uniqueness theorem is present. Accordingly, the derivation chain is self-contained against an external benchmark and no circular step is identified.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 1 invented entities

The central claims lean heavily on domain assumptions: the NAVSIM PDMS/EPDMS metric, the comparability of branches despite non-identical training protocols, the interpretability of the SAE decomposition, and the validity of interpolated candidates. No genuinely new physical or natural entities are proposed; the only new constructs are trained models whose evidence is internal to the paper.

free parameters (4)
  • τ (decisive-win threshold) = 0.2 and 0.5
    The counts of decisive wins (257/253 and 159/153) and the 'near-symmetric long-tail' conclusion depend on this threshold; it is chosen by the authors and not justified independently.
  • γ (DualDriveVLA confidence threshold) = not reported; swept to yield ~15% VLM invocation
    The headline 91.00 PDMS / 15% invocation operating point is selected from the accuracy-compute curve (Fig. 4) by sweeping γ; other thresholds give different trades.
  • α interpolation fractions = 0.1, ..., 0.9 (11 candidates)
    The candidate set in Eq. (3)-(4) is defined by this hand-selected grid on the style axis; the achievable best-of-6 ceiling depends on this grid.
  • SAE dims and cross weight = ds=64, du=16; cross weight=0.1
    Table 1/6 show R2_cross and Δcross vary with the cross weight; the 'non-transferable residual' conclusions use this specific hyperparameter setting.
axioms (5)
  • domain assumption NAVSIM PDMS/EPDMS is a valid and transferable measure of driving quality
    All complementarity gains and selections are measured on this offline pseudo-closed-loop metric (Appendix B); if it does not track real closed-loop performance, the central behavioral claims weaken.
  • domain assumption The two branches are comparable despite different training protocols
    §3.2 asserts fairness, but the VLM branch receives language-domain adaptation and is frozen, while the vision-only branch is trained jointly during IL; differences may be attributable to protocol or capacity, not modality.
  • standard math Linear CKA/CCA and the SAE cross-reconstruction objective are valid representation-similarity diagnostics
    Used as quantitative evidence for a 'shared subspace'; the SAE metric is partly optimized by the training objective, though the shuffled-pair control provides some in-paper validation.
  • domain assumption Linear interpolation of waypoints yields legitimate candidate trajectories
    Eq. (3) creates interpolated candidates with no feasibility or dynamics constraint, so some selection gains may come from trajectory smoothing rather than semantic complementarity.
  • domain assumption The trajectory scorer's predicted PDMS suffices to rank candidates
    §6.1 trains the scorer to predict the same PDMS sub-scores used for evaluation; the reported gain assumes the surrogate ranking transfers to the true metric.
invented entities (1)
  • Shared-Unique Sparse Autoencoder latent decomposition no independent evidence
    purpose: Separates and quantifies shared versus branch-specific representational factors; serves as the main RQ1 diagnostic.
    This is a learned model, not an independently measured phenomenon. Its 'unique subspaces' have no falsifiable handle outside the paper, though the shuffled-pair control provides internal validation.

pith-pipeline@v1.3.0-alltime-deepseek · 20981 in / 16178 out tokens · 168801 ms · 2026-08-04T06:02:26.182806+00:00 · methodology

0 comments
read the original abstract

Vision-Language-Action (VLA) driving augments end-to-end (E2E) planning with language-enabled visual backbones, yet it remains unclear how vision-language models (VLMs) differ from standard vision-only encoders and whether these differences survive policy learning. We study this question under a unified VLM-hidden + diffusion-policy paradigm, comparing multiple VLM families/scales (e.g., InternVL3 and Qwen3VL) with vision-only encoders (e.g., ResNet, ViT, and EVA-CLIP). We ask three questions: how similar are their representations, do residual differences induce meaningful behavioral differences, and how can they improve accuracy-cost trade-offs? We find that VLM and vision-only policies share a substantial common subspace after policy learning, yet both retain non-transferable residual subspaces. Using a Shared-Unique SAE, we show that these residual factors are behaviorally relevant: vision-only encoders are stronger in simple, geometry-dominant scenarios, whereas VLMs are stronger in long-tail, semantically complex, and interaction-heavy cases. The two policy families also exhibit distinct driving styles, with vision-only models being more conservative on average and VLMs more assertive. Exploiting the complementarity between a VLM branch and a ViT branch yields an oracle upper bound of 93.58 PDMS on NAVSIM. We introduce HybridDriveVLA, which runs both branches and uses a learned trajectory scorer for selection, improving PDMS to 92.10 (+1.30 over the VLM baseline), and DualDriveVLA, a fast-slow variant that invokes the VLM in only 15% of scenarios, achieving 91.00 PDMS (+0.20) with about a 1.9x latency speedup over the VLM baseline. Code will be released.

Figures

Figures reproduced from arXiv: 2602.10719 by Bingchuan Sun, Canyu Chen, Cheng Chi, Chenxu Dang, Haiyan Liu, Jason Bao, Sining Ang, Xuanyao Mao, Xuliang, Yan Wang, Yuguang Yang.

Figure 1
Figure 1. Figure 1: Shared vs. model-specific representation geometry after alignment. Each model’s features are aligned to the ResNet￾101 feature space (used as the reference) via (orthogonal) Pro￾crustes alignment (Schonemann ¨ , 1966), then visualized using a 2D PCA projection fitted on the concatenation of all aligned features. For each model, the curves are KDE probability-mass contours enclosing the smallest regions con… view at source ↗
Figure 2
Figure 2. Figure 2: Backbone vs. DiT representation similarity mea￾sured by CKA. We compute pairwise linear CKA between fea￾ture representations from different model branches and visualize the resulting similarity matrices for (left) the visual backbone features and (right) the DiT features. While backbone represen￾tations show high agreement among vision-only encoders (e.g., ViT/ResNet/EVA-CLIP) but low similarity to the VLM… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of our dual-branch RecogDrive system and analysis points. A VLM branch (as in the original RecogDrive) and a vision-only branch (ViT/ResNet/EVA-CLIP) provide alternative visual representations to a diffusion Transformer planner (DiT) and action decoder. The two branches use the same planner architecture but are instantiated as separate policies (no weight sharing), producing two candidate trajecto… view at source ↗
Figure 4
Figure 4. Figure 4: DualDriveVLA accuracy–compute trade-off by varying the confidence threshold γ. The x-axis shows the fraction of scenarios routed to the fast path (ViT-only). The left y-axis reports the overall PDMS score, and the right y-axis reports inference speed (throughput / latency, as measured in our setup). Higher ViT selection ratio increases speed, while the scorer-based fallback preserves performance by invokin… view at source ↗
Figure 5
Figure 5. Figure 5: CCA canonical-correlation spectra (PCA truncation + whitening. A.5. CCA Canonical-correlation Spectra and Original-space Aligned Energy We visualize the PCA-whitened CCA canonical-correlation spectra and report how much original-space feature energy lies in highly aligned CCA directions. Importantly, the “CCA-aligned subspace” here is a post-hoc linear construct and should not be conflated with the learned… view at source ↗
Figure 6
Figure 6. Figure 6: t-SNE of backbone- and DiT-level features colored by whether VLM outperforms ViT for the scenario. The classes are not separable, suggesting intrinsic difficulty for representation-only gating.((Left: backbone-level features; Right: DiT-level features) [PITH_FULL_IMAGE:figures/full_fig_p019_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative case gallery. Each panel contains a front-camera view and a BEV visualization. Red denotes the VLM trajectory, blue denotes the ViT trajectory, and green denotes the human (expert) trajectory used as imitation-learning supervision. Cases 1–5 primarily highlight longitudinal / speed-profile differences, while Cases 6–10 highlight lateral / path and lane-level preference differences. 22 [PITH_FU… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

21 extracted references · 16 linked inside Pith

  1. [1]

    Is a 3d-tokenized llm the key to reliable autonomous driving?arXiv preprint arXiv:2405.18361,

    Bai, Y ., Wu, D., Liu, Y ., Jia, F., Mao, W., Zhang, Z., Zhao, Y ., Shen, J., Wei, X., Wang, T., et al. Is a 3d-tokenized llm the key to reliable autonomous driving?arXiv preprint arXiv:2405.18361,

    Pith/arXiv arXiv

  2. [6]

    ipad: Iterative proposal-centric end-to-end autonomous driving.arXiv preprint arXiv:2505.15111,

    98.3 95.1 81.5 97.410099.8 98.6 96.5 98.3 83.1 ReCogDrive (Li et al., 2025d) 98.3 95.2 87.1 97.5 98.3 99.899.5 96.686.5 83.6 DiffusionDriveV2 (Zou et al., 2025)97.796.688.9 97.2 97.8 99.8 99.2 96.0 91.0 85.5 HybridDriveVLA(ours) 98.692.289.7 98.598.3 99.8 99.396.687.0 85.5 Guo, K., Liu, H., Wu, X., Pan, J., and Lv, C. ipad: Iterative proposal-centric end-...

    Pith/arXiv arXiv 2025

  3. [7]

    Emma: End-to-end multimodal model for autonomous driving

    Hwang, J.-J., Xu, R., Lin, H., Hung, W.-C., Ji, J., Choi, K., Huang, D., He, T., Covington, P., Sapp, B., et al. Emma: End-to-end multimodal model for autonomous driving. arXiv preprint arXiv:2410.23262,

    Pith/arXiv arXiv

  4. [8]

    Senna: Bridging large vision-language models and end-to-end autonomous driving.arXiv preprint arXiv:2410.22313,

    Jiang, B., Chen, S., Liao, B., Zhang, X., Yin, W., Zhang, Q., Huang, C., Liu, W., and Wang, X. Senna: Bridging large vision-language models and end-to-end autonomous driving.arXiv preprint arXiv:2410.22313,

    Pith/arXiv arXiv

  5. [9]

    Driving on registers.arXiv preprint arXiv:2601.05083,

    Kirby, E., Boulch, A., Xu, Y ., Yin, Y ., Puy, G., Zablocki,´E., Bursuc, A., Gidaris, S., Marlet, R., Bartoccioni, F., et al. Driving on registers.arXiv preprint arXiv:2601.05083,

    arXiv

  6. [11]

    Dif- fusiondrive: Truncated diffusion model for end-to-end autonomous driving.arXiv preprint arXiv:2411.15139,

    Liao, B., Chen, S., Yin, H., Jiang, B., Wang, C., Yan, S., Zhang, X., Li, X., Zhang, Y ., Zhang, Q., et al. Dif- fusiondrive: Truncated diffusion model for end-to-end autonomous driving.arXiv preprint arXiv:2411.15139,

    Pith/arXiv arXiv

  7. [12]

    Gpt- driver: Learning to drive with gpt.arXiv preprint arXiv:2310.01415, 2023a

    9 From Representational Complementarity to Dual Systems: Synergizing VLM and Vision-Only Backbones for End-to-End Driving Mao, J., Qian, Y ., Ye, J., Zhao, H., and Wang, Y . Gpt- driver: Learning to drive with gpt.arXiv preprint arXiv:2310.01415, 2023a. Mao, J., Ye, J., Qian, Y ., Pavone, M., and Wang, Y . A language agent for autonomous driving.arXiv pre...

    Pith/arXiv arXiv

  8. [13]

    Sch¨onemann, P

    URL https://www.arxiv.org/abs/2312.03661. Sch¨onemann, P. H. A generalized solution of the orthogonal procrustes problem.Psychometrika, 31(1):1–10,

    Pith/arXiv arXiv

  9. [15]

    Tian, X., Gu, J., Li, B., Liu, Y ., Wang, Y ., Zhao, Z., Zhan, K., Jia, P., Lang, X., and Zhao, H

    URL https://www.arxiv.org/abs/2506.24044. Tian, X., Gu, J., Li, B., Liu, Y ., Wang, Y ., Zhao, Z., Zhan, K., Jia, P., Lang, X., and Zhao, H. Drivevlm: The conver- gence of autonomous driving and large vision-language models.arXiv preprint arXiv:2402.12289,

    Pith/arXiv arXiv

  10. [17]

    Wang, S., Yu, Z., Jiang, X., Lan, S., Shi, M., Chang, N., Kautz, J., Li, Y ., and Alvarez, J

    URL https: //www.arxiv.org/abs/2512.16760v2. Wang, S., Yu, Z., Jiang, X., Lan, S., Shi, M., Chang, N., Kautz, J., Li, Y ., and Alvarez, J. M. Omnidrive: A holis- tic llm-agent framework for autonomous driving with 3d perception, reasoning and planning.arXiv preprint arXiv:2405.01533,

    arXiv

  11. [18]

    Wam-diff: A masked diffusion vla framework with moe and online reinforce- ment learning for autonomous driving.arXiv preprint arXiv:2512.11872,

    Xu, M., Cui, J., Cai, F., Shang, H., Zhu, Z., Luan, S., Xu, Y ., Zhang, N., Li, Y ., Cai, J., et al. Wam-diff: A masked diffusion vla framework with moe and online reinforce- ment learning for autonomous driving.arXiv preprint arXiv:2512.11872,

    arXiv

  12. [19]

    Yuan, C., Zhang, Z., Sun, J., Sun, S., Huang, Z., Lee, C. D. W., Li, D., Han, Y ., Wong, A., Tee, K. P., et al. Drama: An efficient end-to-end motion planner for autonomous driving with mamba.arXiv preprint arXiv:2408.03601,

    Pith/arXiv arXiv

  13. [21]

    arxiv.org/abs/2503.21505v1

    URL https://www. arxiv.org/abs/2503.21505v1. Zecong, T., Zixu, W., Yifei, W., Weitong, L., Tianjian, G., Haoran, L., Tengju, R., Lingyi, M., Zhejun, C., Yichen, Z., Qi, K., Kaixuan, W., and Yu, Z. Auto- dridm: An explainable benchmark for decision-making of vision-language models in autonomous driving.arXiv preprint arXiv:2601.14702v1,

    Pith/arXiv arXiv

  14. [22]

    Zhang, S., Huang, W., Gao, Z., Chen, H., and Lv, C

    URL https:// www.arxiv.org/abs/2601.14702v1. Zhang, S., Huang, W., Gao, Z., Chen, H., and Lv, C. Wisead: Knowledge augmented end-to-end autonomous driving with vision-language model.arXiv preprint arXiv:2412.09951,

    Pith/arXiv arXiv

  15. [23]

    Sce2drivex: A generalized mllm framework for scene-to-drive learning.arXiv preprint arXiv:2502.14917,

    Zhao, R., Yuan, Q., Li, J., Hu, H., Li, Y ., Zheng, C., and Gao, F. Sce2drivex: A generalized mllm framework for scene-to-drive learning.arXiv preprint arXiv:2502.14917,

    Pith/arXiv arXiv

  16. [24]

    Z., Zhang, Y ., Huang, Z., Zhou, B., and Ma, J

    Zhou, Z., Cai, T., Zhao, S. Z., Zhang, Y ., Huang, Z., Zhou, B., and Ma, J. Autovla: A vision-language-action model for end-to-end autonomous driving with adaptive rea- soning and reinforcement fine-tuning.arXiv preprint arXiv:2506.13757,

    Pith/arXiv arXiv

  17. [25]

    Diffusiondrivev2: Rein- forcement learning-constrained truncated diffusion mod- eling in end-to-end autonomous driving.arXiv preprint arXiv:2512.07745,

    Zou, J., Chen, S., Liao, B., Zheng, Z., Song, Y ., Zhang, L., Zhang, Q., Liu, W., and Wang, X. Diffusiondrivev2: Rein- forcement learning-constrained truncated diffusion mod- eling in end-to-end autonomous driving.arXiv preprint arXiv:2512.07745,

    arXiv

  18. [2023]

    Feng, R., Xi, N., Chu, D., Wang, R., Deng, Z., Wang, A., Lu, L., Wang, J., and Huang, Y

    URL https://www.arxiv.org/abs/ 2312.14150v3. Feng, R., Xi, N., Chu, D., Wang, R., Deng, Z., Wang, A., Lu, L., Wang, J., and Huang, Y . Artemis: Autoregressive end-to-end trajectory planning with mixture of experts for autonomous driving.arXiv preprint arXiv:2504.19580,

    Pith/arXiv arXiv

  19. [2024]

    Vadv2: End-to-end vectorized autonomous driving via probabilistic planning

    Chen, S., Jiang, B., Gao, H., Liao, B., Xu, Q., Zhang, Q., Huang, C., Liu, W., and Wang, X. Vadv2: End-to-end vectorized autonomous driving via probabilistic planning. arXiv preprint arXiv:2402.13243, 2024a. Chen, Y ., Wang, Y ., and Zhang, Z. Drivinggpt: Unify- ing driving world modeling and planning with multi- modal autoregressive transformers.arXiv pr...

    Pith/arXiv arXiv

  20. [2025]

    Orion: A holis- tic end-to-end autonomous driving framework by vision- language instructed action generation.arXiv preprint arXiv:2503.19755,

    Fu, H., Zhang, D., Zhao, Z., Cui, J., Liang, D., Zhang, C., Zhang, D., Xie, H., Wang, B., and Bai, X. Orion: A holis- tic end-to-end autonomous driving framework by vision- language instructed action generation.arXiv preprint arXiv:2503.19755,

    Pith/arXiv arXiv

  21. [2026]

    Imagidrive: A unified imagination-and-planning framework for autonomous driving.arXiv preprint arXiv:2508.11428, 2025a

    Li, J., Zhang, B., Jin, X., Deng, J., Zhu, X., and Zhang, L. Imagidrive: A unified imagination-and-planning framework for autonomous driving.arXiv preprint arXiv:2508.11428, 2025a. Li, Y ., Shang, S., Liu, W., Zhan, B., Wang, H., Wang, Y ., Chen, Y ., Wang, X., An, Y ., Tang, C., et al. Drivevla-w0: World models amplify data scaling law in autonomous driv...

    arXiv